NVIDIA releases new physical AI models, plus autonomous vehicle tools

LAS VEGAS — To understand and operate in dynamic environments, physical AI needs to be able to learn, reason, and plan, according to NVIDIA Corp. The company yesterday announced new open models, frameworks, simulation tools, datasets, and artificial intelligence infrastructure for robotics and self-driving vehicle developers.

“The ChatGPT moment for robotics is here,” said Jensen Huang, founder and CEO of NVIDIA. “Breakthroughs in physical AI — models that understand the real world, reason and plan actions — are unlocking entirely new applications.”

“NVIDIA’s full stack of Jetson robotics processors, CUDA, Omniverse, and open physical AI models empowers our global ecosystem of partners to transform industries with AI-driven robotics,” he added.

NVIDA offers models for ‘generalist-specialist’ robots

“Today, most robots are specialists. They are excellent at one single task, but they cannot adapt to anything else,” said Rev Lebaredian, vice president of Omniverse and simulation technology at NVIDIA. “Now, we are seeing generalist robots. Like someone with a bachelor’s degree, they can handle different situations. However, they lack the expert skills for complex jobs.”

“The future belongs to the generalist-specialist,” he asserted. “Think of them as the Ph.D.s of the robot world. They combine broad knowledge with deep expertise, making them versatile and reliable. Building these advanced robots requires an open development platform. Perception alone is not enough, which is why we’re offering the new Cosmos models for reasoning and advanced foundation models with world generation and understanding.”

Today’s machines are single-task and hard to program, said NVIDIA. Making them more capable typically requires enormous capital and expertise to build foundation models, but the Santa Clara, Calif.-based company claimed that its open models allow developers to bypass resource-intensive pretraining.

The new models, all available on Hugging Face, include:

• NVIDIA Cosmos Transfer 2.5 and NVIDIA Cosmos Predict 2.5: Fully customizable world models that enable physically based synthetic data generation and robot policy evaluation in simulation for physical AI
• NVIDIA Cosmos Reason 2: A reasoning vision language model (VLM) that NVIDIA said enables machines to see, understand, and act in the physical world like humans
• NVIDIA Isaac GR00T N1.6: A reasoning vision-language-action (VLA) model, designed to unlock full body control for humanoids so they can move and handle objects simultaneously; it uses NVIDIA Cosmos Reason for reasoning and contextual understanding

---

---

Simulation, compute frameworks designed for robotics

Scalable simulation is essential for training and evaluating robots, but current workflows remain fragmented and difficult to manage, said NVIDIA. Benchmarking is often manual and hard to scale, while end-to-end pipelines require complex orchestration across disparate compute resources.

The company today released new open-source frameworks on GitHub to simplify these pipelines and accelerate the transition from research to real-world use cases.

NVIDIA Isaac Lab-Arena provides a system for robot policy evaluation and benchmarking in simulation, with the evaluation and task layers designed with Lightwheel. The company said it standardizes testing and ensures that robot skills are robust and reliable before they are deployed to physical hardware.

“Isaac Lab-Arena is the world’s first collaborative system for large-scale robot policy evaluation and benchmarking to address this critical gap,” said Lebaredian. “It unifies assets, tasks, training scripts, and the most important robotics community benchmarks, such as Libero and Robocasa. As the community’s single source of truth, Isaac Lab-Arena offers the scaffolding needed to benchmark skills before real-world release.”

NVIDIA OSMO is a cloud-native orchestration framework that unifies robotic development into a single command center. The company said it lets developers define and run workflows such as synthetic data generation, model training, and software-in-the-loop testing.

OSMO works across different compute environments — from workstations to mixed cloud instances — speeding up development cycles, NVIDIA said. OSMO is now available and used by robot developers such as Hexagon, and it is integrated into the Microsoft Azure Robotics Accelerator toolchain.

NVIDIA, Hugging Face team to speed physical AI development

“NVIDIA and Hugging Face are teaming up to unite our communities, connecting 2 million NVIDIA robotic experts with 13 million Hugging Face AI builders,” said Lebaredian. “NVIDIA’s Isaac and GR00T technologies are now built into Hugging Face LeRobot library. This gives developers instant access to models like GR00T N1.6 and simulation frameworks like Isaac Labyrinth for evaluating robot skills.

“On the hardware side, everything just works,” he said. “The open-source Reachy 2 humanoid will run seamlessly on NVIDIA Jetson Thor, letting developers deploy advanced AI models right on the robot. And for desktop projects, the Reachy Mini pairs with DGX Spark to run custom AI, voice, and vision models locally.”

Companies across industries use GR00T

Several companies are already using GR00T-enabled workflows to simulate, train and validate new behaviors for their robots. LEM Surgical uses NVIDIA Isaac for Healthcare and Cosmos Transfer to train the autonomous arms of its Dynamis surgical robot, powered by NVIDIA Jetson AGX Thor and Holoscan.

Several exhibitors at CES this week are using Jetson Thor to meet the demand for humanoid robots with reasoning. Richtech Robotics is launching Dex, a mobile manipulator for complex industrial environments, while RLWRLD has integrated Thor enhance navigation and manipulation of its household robot.

“Boston Dynamics is launching a new Atlas humanoid running on Jetson Thor and trained in Isaac Lab Arena,” said Lebaredian. “Franka Robotics is using the GR00T N model to power its dual-arm manipulator robot.”

“LG Electronics is unveiling a new robot that handles various household chores,” he added. “NEURA is launching a Porsche-designed humanoid powered by GR00T N and developed with Isaac Lab.”

In addition, Humanoid is using GR00T, and XRlabs is using Thor and Isaac for Healthcare to enable surgical scopes to guide surgeons with real-time AI analysis. AGIBOT is introducing systems for both industrial and consumer sectors, as well as Genie Sim 3.0, a robot simulation platform integrated with Isaac Sim.

Salesforce is using Agentforce, Cosmos Reason, and the NVIDIA Blueprint for video search and summarization to analyze video footage captured by its robots and halve incident-resolution times.

NVIDIA brings Blackwell architecture to the industrial edge

NVIDIA said its new Jetson T4000 module is designed to be an affordable, high-performance upgrade for Orin customers, bringing the Blackwell architecture to robotics at $1,999 per 1,000 units. It delivers quadruple the performance of the previous generation of onboard processors, with 2,070 FP4 TFLOPS (trillion floating-point operations per second) and 64GB of memory, all within a configurable 70-watt envelope ideal for energy-constrained autonomy.

“It fits the exact same slot as the T5000, so swapping it in for production is a breeze,” said Lebaredian. “We imagine these modules to power many types of robots, from manipulators to Mars to humanoids.”

NVIDIA said IGX Thor, which will be available later this month, “extends robotics to the industrial edge, offering high-performance AI computing with enterprise software support and functional safety.” Archer is using IGX Thor to advance capabilities in aircraft safety, airspace integration, and autonomy-ready systems.

Partners including AAEON, Advantech, ADLINK, Aetina, AVerMedia, Connect Tech, EverFocus, ForeCR, Lanner, RealTimes, Syslogic, Vecow, and YUAN offer Thor-powered systems equipped for edge AI, robotics, and embedded applications.

In addition, Caterpillar is expanding its collaboration with NVIDIA for equipment using AI in construction and mining.

Alpamayo models and tools intended for reasoning-based AVs

NVIDIA today also released an opening reasoning VLA model for “long-tail” driving challenges that have been traditionally addressed with separate perception and planning. Its new Aplamayo family includes simulation tools and datasets for autonomous vehicle (AV) development.

Alpamayo 1, AlpaSim, and physical AI open datasets enable the development of models and vehicles that perceive, reason and act with humanlike judgment for greater safety, robustness, and scalability, NVIDIA asserted.

“Autonomous driving is the first real example of physical AI,” said Ali Kani, vice president and general manager for NVIDIA Automotive. “The software in cars has gone through a lot of change over the last 15 years. The first gen of AV was perception-only. The car could see the world, but driving logic was mostly hand-coded or rules-based.”

“Step 1.5 added model-based planning on top of perception, giving more structured behavior to the experience,” he explained. “Step 2 used generative end-to-end AI that learned driving behavior directly from data. And we’re now moving to AV3, agentic physical AI, where the vehicle reasons, plans, and drives like a capable assistant in the real world.”

With Alpamayo, mobility companies such as Jaguar Land Rover, Lucid, and Uber, as well as researchers such as Berkeley DeepDrive, can accelerate reasoning‑based SAE Level 4 deployment roadmaps. said NVIDIA.

The post NVIDIA releases new physical AI models, plus autonomous vehicle tools appeared first on The Robot Report.